University of Houston

We developed an automated skin cancer detection system. First, we proposed a Narrow Band Graph Partitioning Active Contour(NBGPAC) segmentation method. We validated the method on 211 skin lesion images with ground truth obtained from experienced dermatologists. Our method achieves an F-score of 91%, outperforming two other popular active contours methods Graph Partitioning Active Contour(GPAC) and Chan-Vese model with F-scores 86% and 85% respectively. Second, we employed kernel method for melanoma detection. Similarity between two lesions with bag-of-feature representation is evaluated by χ² kernel, diffusion kernel, and pyramid match kernel. Combining kernels computed from color, SIFT, and wavelet feature, this whole system achieves 89% area under ROC, comparable to state-of-art automated systems MelaFind and SolarScan.

Our proposed future works include identifying specific texture patterns inside skin lesion with multiple instance learning and semi-supervised learning. Active sampling will also be investigated to reduce workload of physicians. Preliminary result with an active learning strategy indicates that we can reduce labeling effort of domain expert by half while maintain a similar performance. Applications on other medical images will also be our future work.